Large AI models: Pioneering innovations in accurate weather forecasting

<p indent="0mm">Weather forecasting not only stands at the forefront of international scientific inquiry but also holds significant economic and societal value. The paradigm of numerical weather prediction (NWP), dating back to the 1950s, relies on solving partial differential equations that describe atmospheric motions to predict future atmospheric states, often requiring hundreds of core hours on supercomputers. The European Centre for Medium-Range Weather Forecasts (ECMWF) and its Integrated Forecasting System (IFS) have notably excelled among various national and institutional operational forecast models, guiding the direction of NWP developments. The World Meteorological Organization heralds the NWP revolution as one of the most important scientific, technological, and social advancements of the 20th century. However, traditional methods for enhancing forecasting accuracy have reached a plateau, while the emergence of AI technologies offers new avenues to overcome these limitations. In recent years, with the accumulation of high-quality meteorological data, artificial intelligence (AI), particularly deep learning techniques, has been increasingly employed to model diverse Earth system processes across all spatiotemporal scales. The intersection of AI with traditional data assimilation and ensemble forecasting methods is deepening. Unlike traditional approaches, data-driven AI meteorological models learn underlying physical laws directly from extensive data sets without relying on partial differential equations. A prime example of this in 2023 is Huawei’s PanGu Weather AI model, which has demonstrated accuracy comparable to the IFS at significantly lower computational costs. This model’s success was highlighted in a research paper published in Nature on July 20, 2023. The achievements of data-driven modeling have made AI meteorological models one of the most notable breakthroughs in the “2023 Top Ten Scientific Advances in China”. Artificial Intelligence (AI) meteorological models have garnered significant attention in the “AI + Meteorology” interdisciplinary field by leveraging highly nonlinear neural network architectures to directly learn and represent physical processes from vast data sets. These data-driven models achieve or even surpass traditional weather forecasting methods without relying on explicit physical equations, thus sparking a wave of “AI + Meteorology” modeling initiatives. Apart from the notable PanGu Weather, several independently developed AI atmospheric and ocean models have emerged in China, including FuXi, FengWu, TianXing, NowCastNet, AI-GOMS, and XiHe. Institutions such as Fudan University, Tongji University, Tsinghua University, National University of Defense Technology, Shanghai AI Lab, and the National Meteorological Center are enhancing these AI models’ forecasting capabilities, for instance, by improving spatio-temporal resolutions and providing ensemble forecasts. Some models incorporate existing physical laws to a degree, offering better physical interpretability and forecasting skills than traditional numerical models. Internationally, research institutions are advancing AI meteorological model development, with projects like Google DeepMind’s GenCast, Google’s MetNet series and NeuralGCM, Microsoft’s ClimaX, and ECMWF’s AIFS. These efforts are predominantly data-driven and exceed the numerical forecasting techniques in precision. Global meteorological agencies including the China Meteorological Administration and ECMWF have commenced real-time performance evaluations of multiple AI models like PanGu Weather, FuXi, GraphCast, and FourCastNet, aiming to integrate these models into operational forecasting. These developments vividly demonstrate the profound impact of cutting-edge “AI + Meteorology” research and hold the promise of delivering more accurate forecasts at lower costs. Current evaluations indicate substantial potential in AI-driven large-scale meteorological models, warranting anticipation for their future advancements. This paper reviews the current research progress on large-scale weather models both domestically and internationally, discusses their scientific significance, and explores future prospects for these models.